Strip mining

Strip mining is a fundamental +O3 transformation. Used by itself, strip mining is not profitable. However, it is used by loop blocking, loop unroll and jam, and, in a sense, by parallelization.

Strip mining involves splitting a single loop into a nested loop. The resulting inner loop iterates over a section or strip of the original loop, and the new outer loop runs the inner loop enough times to cover all the strips, achieving the necessary total number of iterations. The number of iterations of the inner loop is known as the loop’s strip length.

DO I = 1, 10000
  A(I) = A(I) * B(I)
ENDDO

Strip mining this loop using a strip length of 1000 yields the following loop nest:

DO IOUTER = 1, 10000, 1000
  DO ISTRIP = IOUTER, IOUTER+999
    A(ISTRIP) = A(ISTRIP) * B(ISTRIP)
  ENDDO
ENDDO

In this loop, the strip length integrally divides the number of iterations, so the loop is evenly split up. If the iteration count was not an integral multiple of the strip length—if I went from 1 to 10500 rather than 1 to 10000, for example—the final iteration of the strip loop would execute 500 iterations instead of 1000.